Keynote: AutoML for TinyML

It is widely anticipated that machine inference will be a key driver for billions of intelligent edge devices. The de facto enabler for machine intelligence is accurate and efficient Deep Neural Networks(DNN). As DNNs state-of-the-art DNNs are typically large and compute-heavy, this has promoted research in AutoML for generating tiny DNN models. AutoML frameworks aid in the automated searching of tiny neural architectures and the corresponding deployment on TinyML platforms. The mainstream AutoML research is focused on the fast search of DNN architectures, to surpass the performance of human-designed ones. However, in many scenarios, the primary goal is not to search for the most accurate model, but the one that is reasonably accurate, and most importantly, efficient. This efficiency includes the model's capability to meet platform-specific constraints and minimization of one or more objectives, e.g. power, latency, memory, etc. Towards this, the current talk outlines typical Neural Architecture Search (NAS) / AutoML components for TinyML applications. A primer on this topic and the customization of a Reinforcement Learning AutoML is already available in http://tinyurl.com/25b2t82u. The theme of these talks is the same - an algorithmic perspective of AutoML, which resonates well with software developers. In the current talk, I focus on search space design for various TinyML applications using a genetic algorithmic framework.